Building an AI Writing System That Works

Faced with high cost for content creation challenge, I developed an AI-powered writing system that generates SEO-optimized articles at scale. By engineering sophisticated AI prompts and building Python scripts.

Client

internal project

Timeline

December to March

Services

prompt engineering

I'm a UX designer, not a programmer. But I've learned that the best solutions often come from unexpected places. I decided to build something that didn't exist yet.

Challenge

Picture this: You have 500 keywords sitting in a spreadsheet. Each one represents a potential customer searching for answers. Each one could drive traffic to your business. But creating quality content for 500 keywords? That's 500 articles. At hourly rates of over 1000 CZK, we would look at well over tens of thousands of Czech crowns for all these articles. However, what if we generate all the articles with AI? We could build a baseline that covers all the necessary keywords in our domain and then improve the already existing articles while already being present in searches through AI.

Exploration Study - Comprehending the Human Component

Before diving into technical solutions, I collaborated with our stakeholder and design lead to define success metrics and scope. We agreed on a "test fast, learn fast" approach—deploying to real traffic rather than lengthy pre-testing phases.

This alignment proved crucial when challenges emerged later in the process.

Our initial idea was to create a chain of command in tools that let you connect different tasks, such as Make or N8N. However, they are rather expensive services, and because the process was iterative, meaning I was reworking one sequence multiple times, I have decided that using AI and Python would be the best approach, at least for the MVP version of our project.

I crafted detailed prompts using the CO-STAR framework. These weren't simple instructions—they were comprehensive blueprints that taught Claude AI our writing style, tone preferences, and content structure requirements.

The breakthrough came when I realized AI needed examples of what NOT to do, not just what to do. I fed it our best content as reference material and our brand guidelines as gospel.

My Role

I worked as the lead on this project. I managed to research the technical solutions, then I worked on crafting initial prompts to understand how to work with AI on such a complex task. I have worked with AI on creating sequences of scripts in Python to create the articles for our keywords and finally uploading them onto our webpage.

Building solutions

Research Insights & Opportunities

My Learnings

Prompt engineering is everything. The difference between generic AI content and professional-quality articles lies in detailed, specific prompts with clear examples and guidelines.

Technical collaboration accelerates innovation. While learning to code independently was valuable, involving technical specialists earlier would have solved challenges faster and more efficiently.

Challenges on the Way

Methods

I worked as the lead on this project. I managed to research the technical solutions, then I worked on crafting initial prompts to understand how to work with AI on such a complex task. I have worked with AI on creating sequences of scripts in Python to create the articles for our keywords and finally uploading them onto our webpage.

Goals

  • Voice consistency: Every article needed to sound like us

  • SEO optimization: Keywords placed naturally, proper structure

  • Educational value: Real information that helps readers

  • Technical integration: Ready-to-publish HTML for our CMS

Methods

The final workflow is elegant in its simplicity:

  1. Input: Keyword list and reference materials

  2. Processing: AI generates multiple article versions

  3. Validation: Automated quality checks score the articles to ensure standards

  4. Selection: The Algorithm chooses the best version on the highest score and combines strneghts from other articles into it.

  5. Output: Ready-to-publish HTML formatted for our CMS

Summary

We challenged conventional SEO wisdom that says "focus on high-volume, low-competition keywords." Instead, we covered everything. The data suggests this comprehensive approach works—our traffic patterns show consistent engagement across diverse content pieces.

Next phase development includes:

  • Vector database integration: Building a comprehensive knowledge base that AI can reference for more accurate, company-specific content

  • Modular script architecture: Breaking the generation process into smaller, specialized components that work independently before combining results

  • Incremental AI processing: Rather than generating complete articles in one pass, the system will create sections separately, then intelligently combine them into cohesive pieces

Expected Impact

Our goal was to build a solution where we would only plug in the words for which we wanted an article created, and have AI take care of the rest. For MVP, this has proven to be too ambitious; however, we have laid solid ground for the next steps to follow.

Outcome


The numbers tell the story better than I ever could:

Content Production:

  • Generated articles for 100% of target keywords

  • Reduced content creation time by 95%

  • Maintained consistent brand voice across all pieces

  • Created ready-to-publish content at scale

Traffic Performance (Based on Analytics):

  • Steady organic traffic growth over 2 months

  • Peak traffic reaching 2,000+ page views

  • Consistent user engagement with generated content

  • Multiple traffic spikes indicate content resonance

Business Impact:

  • Eliminated tens of thousands of czech crowns in copywriting costs

  • Covered the competitive keyword landscape completely

  • Enabled rapid content expansion for new topics

  • Built internal capability for ongoing content needs

My Learnings

Research Insights & Opportunities

Prompt engineering is everything. The difference between generic AI content and professional-quality articles lies in detailed, specific prompts with clear examples and guidelines.

Technical collaboration accelerates innovation. While learning to code independently was valuable, involving technical specialists earlier would have solved challenges faster and more efficiently.